Compatibility patch

[0.8.2] - 2023-06-30

Changed

• Added GATE backbone for Tabular integrations (#1559)

Fixed

• Fixed datamodule can't load files with square brackets in names (#1501)
• Fixed channel dim selection on segmentation target (#1509)
• Fixed used of jsonargparse avoiding reliance on non-public internal logic (#1620)
• Compatibility with pytorch-tabular>=1.0 (#1545)
• Compatibility latest numpy (#1595)

New Contributors

• @kjappelbaum made their first contribution in #1503
• @yurijmikhalevich made their first contribution in #1501
• @pirj made their first contribution in #1520
• @ArjunSharda made their first contribution in #1517
• @izikgo made their first contribution in #1509
• @manujosephv made their first contribution in #1559
• @mauvilsa made their first contribution in #1620

Full Changelog: 0.8.1.post0...0.8.2

Dependency's adjustments

What's Changed

• fixed type of 'n_gram' from bool to int in TranslationTask by @BrightXiaoHan in #1486
• pinned torchmetrics version for compatibility by @Borda in #1495
• pinned sahi to fix object detection when installing in a fresh environment by @ethanwharris in #1496
• pinned numpy for type compatibility by @Borda in #1504

New Contributors

• @BrightXiaoHan made their first contribution in #1486
• @kjappelbaum made their first contribution in #1503

Full Changelog: 0.8.1...0.8.1.post0

Minor compatibility patch

What's Changed

• Add CLIP backbones for text / image classification by @ethanwharris in #1458
• Replace DP/DDP/DDPSpawn plugins to strategies, keep the old for compatibility by @krshrimali in #1451
• Integration of lightning_utilties function into flash by @uakarsh in #1457
• refactored image_classifier_head to classifier_head by @Abelarm in #1464
• Raise better error if icevision not installed if module isn't found (loading data) by @krshrimali in #1474
• Add support for Lightning 1.8 + Fixes for the CI by @krshrimali in #1470 and #1479
• Fix compatibility with TM 0.10 by @ethanwharris in #1469

New Contributors

• @Abelarm made their first contribution in #1464

Full Changelog: 0.8.0...0.8.1

TPU Support, Remote Data Loading, Video Classification from tensors: Feature Rich Release

We are elated to announce the release of Lightning Flash v0.8, a feature-rich release with improved testing to ensure better user experience for all our lovely users! The team at Lightning AI and our community contributors have been working hard for this release, and nothing makes us happier to share all their lovely contributions with you.

We discuss major features and changes below. For a curated list, scroll to the bottom to see all the pull requests included for this release.

TPU Support 🦸🏻

Before this release, Lightning Flash worked well on a single-core TPU (training, validation, and prediction), but failed comprehensively on multiple cores. This release has enabled training and validation support for multi-core TPUs, allowing users to try out their models on TPUs using Lightning Flash. Prediction of multi-core TPUs is an ongoing effort, and we hope to bring it to you in the near future.

	Before v0.8	After v0.8
Single core	Training, Validation, Prediction	Training, Validation, Prediction
Multiple cores	Not supported	Training, Validation

As we move ahead, and we see more users trying the TPUs with Lightning Flash, we expect that there might be unseen errors or issues, and we will be looking forward to addressing them as we get a chance. So please don't hesitate to let us know your experience!

Remote Data Loading: fsspec arrives into Lightning Flash ☁️

Before this release, users had to download a dataset or a file from the URL and pass it to our data loader classes. This was a pain point that we are happy to let go of in this release. Starting v0.8, you'll not have to download any of those files locally, and you can just pass the file URL - and expect it to work!

	Before v0.8	After v0.8
Example	Download titanic.csv from the URL and pass the path to the train_file argument: from flash.tabular import TabularClassificationData datamodule = TabularClassificationData.from_csv( categorical_fields=["Age", "Cabin"], numerical_fields="Fare", target_fields="Survived", train_file="titanic.csv", val_split=0.1, batch_size=8, )	Just pass the URL to train_file argument: from flash.tabular import TabularClassificationData datamodule = TabularClassificationData.from_csv( categorical_fields=["Age", "Cabin"], numerical_fields="Fare", target_fields="Survived", train_file="https://pl-flash-data.s3.amazonaws.com/titanic.csv", val_split=0.1, batch_size=8, )

For more details, feel free to check out the documentation here.

Video Classification from Tensors 📹

At times, it's required to load raw data, or pre-process videos before progressing to loading data and training the model. These raw data for Video Classification, are mostly available as tensors, and before this release - one had to save them again in video files, and pass the paths to the data loading classes in Flash. Starting this release, we now support loading data from tensors for Video Classification.

import torch
from flash.video import VideoClassifier, VideoClassificationData
import flash

# 5 number of frames, 3 channels, height = 10 and width = 10
mock_tensors = torch.randint(size=(3, 5, 10, 10), low=0, high=255)
datamodule = VideoClassificationData.from_tensors(
    train_data=[mock_tensors, mock_tensors],  # can also stack: torch.stack((mock_tensors, mock_tensors))
    train_targets=["patient", "doctor"],
    predict_data=[mock_tensors],
    batch_size=1,
)

model = VideoClassifier(num_classes=datamodule.num_classes, pretrained=False, backbone="slow_r50", labels=datamodule.labels)
trainer = flash.Trainer(max_epochs=1)
trainer.finetune(model, datamodule=datamodule)

This will also come in handy for those having multi-modal pipelines who don't want to save the output of a model to files and instead pass the raw data to the next model, saving you quite a lot of time wasted in the conversion process.

Refactored Transforms in Lightning Flash ⚙️

One of the community-driven contributions that we are proud to share. Before this release, a user had to pass an input transform class for each stage, which was cumbersome. With this release, you can just pass transform=<YourTransformClass> to the required method. This is a breaking change, and if you are not sure how to resolve this, please create an issue and we'll be happy to help!

	Before v0.8	After v0.8
Example	dm = XYZTask_DataModule.from_xyz( train_file=train_file, val_file=val_file, test_file=test_file, predict_file=predict_file, train_transform=InputTransform, val_transform=InputTransform, test_transform=InputTransform, predict_transform=InputTransform, transform_kwargs=transform_kwargs, )	dm = XYZTask_DataModule.from_xyz( train_file=train_file, val_file=val_file, test_file=test_file, predict_file=predict_file, transform=InputTransform(**transform_kwargs), )

Note that, within your InputTransform class, you can have <stage>_per_batch_transform_on_device methods to support various stages.

class SampleInputTransform(InputTransform):
    def per_sample_transform(self):
        def fn(x):
            return x
       return fn

    def train_per_batch_transform_on_device(self) -> Callable:
        return ...

    def val_per_batch_transform_on_device(self) -> Callable:
        return ...

    def test_per_batch_transform_on_device(self) -> Callable:
        return ...

    def predict_per_batch_transform_on_device(self) -> Callable:
        return ...

Object Detection in Flash is now servable 💁

If you aren't aware yet, Lightning Flash supports serving models. Starting this release, Object Detection is added to the beautiful category of tasks that can be served using Lightning Flash. Below is an example of how the inference server code for object detection will look like:

# Inference Server
from flash.image import ObjectDetector

model = ObjectDetector.load_from_checkpoint("https://flash-weights.s3.amazonaws.com/0.8.0/object_detection_model.pt")
model.serve()

For more details, check out the documentation here.

Added

• Added support for from_tensors for VideoClassification (#1389)
• Added fine tuning strategies for DeepSpeed (with parameter loading and storing omitted) (#1377)
• Added torchvision as a requirement to datatype_audio.txt as it's used for Audio Classification (#1425)
• Added figsize and limit_nb_samples for showing batch images (#1381)
• Added support for from_lists for Tabular Classification and Regression (#1337)
• Added support for from_dicts for Tabular Classification and Regression (#1331)
• Added support for using the ImageEmbedder SSL training for all image classifier backbones (#1264)
• Added support for audio file formats to AudioClassificationData (#1085)
• Added support for Flash serve to the ObjectDetector (#1370)
• Added support for loading ImageClassificationData from PIL images with from_images (#1372)
• Added support for loading ObjectDetectionData with from_numpy, from_images, and from_tensors (#1372)
• Added support for remote data loading with fsspec (#1387)
• Added support for TSV files to from_csv methods (#1387)
• Added support for more formats when loading audio files (#1387)
• Added support to use any task as an embedder by calling as_embedder (#1396)
• Added support for normalization of images in SemanticSegmentationData (#1399)

Changed

• Changed the ImageEmbedder dependency on VISSL to optional (#1276)
• Changed the transforms in SemanticSegmentationData to use albumentations instead of Kornia (#1313)

Removed

• Removed support for audio files with sd2 extension, because SoundFile (for sd2 extension) doesn't accept fsspec objects. ([#1409](https://github.com/Lightning-AI/lightn...

Bi-Weekly Patch Release

[0.7.5] - 2022-05-11

Fixed

• Fixed image classification data show_train_batch for subplots with rows > 1. (#1315)
• Fixed support for all the versions (including the latest and older) of baal. (#1315)
• Fixed a bug where a loaded TabularClassifier or TabularRegressor checkpoint could not be served (#1324)
• Fixed a bug where the freeze_unfreeze and unfreeze_milestones finetuning strategies could not be used in tandem with a onecyclelr LR scheduler (#1329)
• Fixed a bug where the backbone learning rate would be divided by 10 when unfrozen if using the freeze_unfreeze or unfreeze_milestones strategies (#1329)

Contributors

@Borda @ethanwharris @kaushikb11 @krshrimali

If we forgot someone let us know 😃

Bi-Weekly Patch Release

[0.7.4] - 2022-04-27

Fixed

• Fixed a bug where LR schedulers from HuggingFace could not be used with newer versions of PyTorch Lightning (#1307)
• Fixed a bug where the default Flash zero configurations for ObjectDetector, InstanceSegmentation, and KeypointDetector would error with the latest version of some requirements (#1306)
• Fixed plain LightningModule support for Flash data modules. (#1281)

Contributors

@Borda @ethanwharris @krshrimali @rohitgr7

If we forgot someone let us know 😃

Bi-Weekly Patch Release

[0.7.3] - 2022-04-13

Fixed

• Fixed a bug where some backbones were incorrectly listed as available for the ObjectDetector, InstanceSegmentation, and KeypointDetector (#1267)
• Fixed a bug where the backbone would not be frozen when finetuning the SpeechRecognition task (#1275)
• Fixed a bug where the backbone would not be frozen when finetuning the QuestionAnswering task with certain model types (#1275)

Contributors

@ethanwharris @krshrimali

If we forgot someone let us know 😃

Bi-Weekly Patch Release

[0.7.2] - 2022-03-30

Fixed

• Fixed examples (question answering), where NLTK's punkt module needs to be downloaded first. (#1215)
• Fixed normalizing inputs to video classification (#1213)
• Fixed a bug where pretraining_transforms in the ImageEmbedder was never called. (1196)
• Fixed a bug where BASE_MODEL_NAME was not in the dict for dino and moco strategies. (1196)
• Fixed support for torch==1.11.0 (#1234)
• Fixed DDP spawn support for ObjectDetector, InstanceSegmentation, and KeypointDetector (#1222)
• Fixed a bug where InstanceSegmentation would fail if samples had an inconsistent number of bboxes, labels, and masks (these will now be treated as negative samples) (#1222)
• Fixed a bug where collate functions were never called in the ImageEmbedder class. (#1217)
• Fixed a bug where ObjectDetector, InstanceSegmentation, and KeypointDetector would log train and validation metrics with the same name (#1252)
• Fixed a bug where using ReduceLROnPlateau would raise an error (#1251)
• Fixed GPU support for self-supervised training with the ImageEmbedder (#1256)

Contributors

@aisensiy @andife @aniketmaurya @Borda @dudeperf3ct @ethanwharris @krshrimali

If we forgot someone let us know 😃

Bi-Weekly Patch Release

[0.7.1] - 2022-03-01

Added

• Added the normalization parameters of torchvision.transforms.Normalize as transform_kwargs in the ImageClassificationInputTransform (#1178)
• Added available_outputs method to the Task (#1206)

Fixed

• Fixed a bug where DDP would not work with Flash tasks (#1182)
• Fixed DDP support for VideoClassifier (#1189)
• Fixed a bug where buffers in loss functions were not correctly registered in the Task (#1203)
• Fixed support for passing a sampler instance to from_* methods / the DataModule (#1204)

Contributors

@aisensiy @AndresAlgaba @Borda @ethanwharris

If we forgot someone due to not matching commit email with GitHub account, let us know :]

PyTorch Tabular, Enhanced Data Loading and Stability

[0.7.0] - 2022-02-15

Added

• Added support for multi-label, space delimited, targets (#1076)
• Added support for tabular classification / regression backbones from PyTorch Tabular (#1098)
• Added Flash zero support for tabular regression (#1098)
• Added support for COCO annotations with non-default keypoint labels to KeypointDetectionData.from_coco (#1102)
• Added support for from_csv and from_data_frame to VideoClassificationData (#1117)
• Added support for SemanticSegmentationData.from_folders where mask files have different extensions to the image files (#1130)
• Added FlashRegistry of Available Heads for flash.image.ImageClassifier (#1152)
• Added support for ObjectDetectionData.from_files (#1154)
• Added support for passing the Output object (or a string e.g. "labels") to the flash.Trainer.predict method (#1157)
• Added support for passing the TargetFormatter object to from_* methods for classification to override target handling (#1171)

Changed

• Changed Wav2Vec2Processor to AutoProcessor and seperate it from backbone [optional] (#1075)
• Renamed ClassificationInput to ClassificationInputMixin (#1116)
• Changed the default learning_rate for all tasks to be None, corresponding to the default for your chosen optimizer (#1172)

Fixed

• Fixed a bug when not explicitly passing embedding_sizes to the TabularClassifier and TabularRegressor tasks (#1067)
• Fixed a bug where under some circumstances transforms would not get called (#1072)
• Fixed a bug where prediction would sometimes give the wrong number of outputs (#1077)
• Fixed a bug where passing the val_split to the DataModule would not have the desired effect (#1079)
• Fixed a bug where passing predict_data_frame to ImageClassificationData.from_data_frame raised an error (#1088)
• Fixed a bug where segmentation files / masks were loaded with an inconsistent ordering (#1094)
• Fixed a bug with AudioClassificationData.from_numpy (#1096)
• Fixed a bug when using SpeechRecognitionData.from_files for training / validating / testing (#1097)
• Fixed a bug when using SpeechRecognitionData.from_csv or from_json when predicting without targets (#1097)
• Fixed a bug where SpeechRecognitionData.from_datasets did not work as expected (#1097)
• Fixed a bug where loading data for prediction with SemanticSegmentationData.from_folders raised an error (#1101)
• Fixed a bug when passing a predict_folder argument to from_coco / from_voc / from_via in IceVision tasks (#1102)
• Fixed ObjectDetectionData.from_voc and ObjectDetectionData.from_via (#1102)
• Fixed a bug where InstanceSegmentationData.from_coco would raise an error if not using file-based masks (#1102)
• Fixed InstanceSegmentationData.from_voc (#1102)
• Fixed a bug when loading tabular data for prediction without a target field / column (#1114)
• Fixed a bug when loading prediction data for graph classification without targets (#1121)
• Fixed a bug where loading Seq2Seq data for prediction would not work if the target field was not present (#1128)
• Fixed a bug where from_fiftyone classmethods did not work correctly with a predict_dataset (#1136)
• Fixed a bug where the labels property would return None when using ObjectDetectionData.from_fiftyone (#1136)
• Fixed a bug where TabularData would not work correctly with no categorical variables (#1144)
• Fixed a bug where loading TabularForecastingData for prediction would only yield a single sample per series (#1149)
• Fixed a bug where backbones for the ObjectDetector, KeypointDetector, and InstanceSegmentation tasks were not always frozen correctly when finetuning (#1163)
• Fixed a bug where DataModule.multi_label would sometimes be None when it had been inferred to be False (#1165)

Removed

• Removed the Seq2SeqData base class (use TranslationData or SummarizationData directly) (#1128)
• Removed the ability to attach the Output object directly to the model (#1157)

Contributors

@Actis92 @ajinkyaindulkar @bartonp2 @Borda @daMichaelB @ethanwharris @flozi00 @karthikrangasai @MikeTrizna

If we forgot someone due to not matching commit email with GitHub account, let us know :]